PaperlayerView exact revision on arXiv

Source-linked AI summary

CLASH: Evaluating Language Models on Judging High-Stakes Dilemmas from Multiple Perspectives

Ayoung Lee, Ryan Sungmo Kwon, Peter Railton, Lu Wang

arXiv:2504.10823v4cs.CLcs.AI

TL;DR

High-stakes value dilemmas require models to reason across conflicting, contextualized perspectives, but prior benchmarks largely focused on everyday scenarios. CLASH provides 345 dilemmas and 3,795 character perspectives to evaluate ambivalence, discomfort, value shifts, reasoning behavior, and steerability. Models struggle especially with ambivalent decisions and shifted values, while steerability varies with model preferences and framing.

  • Problem

    Existing work provides limited evidence about whether LLMs can make appropriate judgments across pluralistic values in high-stakes dilemmas.

    Source: §1 · p. 1

  • Method

    CLASH benchmarks high-stakes dilemmas using contextualized character perspectives and questions about acceptability, discomfort, value shifts, and conditional steerability.

    Source: §1 · pp. 1–2, §2 · pp. 3–4

  • Results

    Models struggle with ambivalent decisions and shifted values, while conditional steerability is negatively correlated with preferences and generally improves under third-party framing.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Value reasoning exhibits failure patterns distinct from math and game reasoning, including early commitment and overcommitment, and exposes limitations in frontier LLM judgment.

    Source: §6 · p. 10

    CLASH is a methodological and diagnostic benchmark, not evidence supporting autonomous deployment of LLMs as high-stakes decision-makers.

    Source: ETHICS STATEMENT · pp. 10–11

Abstract

from arXiv · show

Navigating dilemmas involving conflicting values is challenging even for humans in high-stakes domains, let alone for AI, yet prior work has been limited to everyday scenarios. To close this gap, we introduce CLASH (Character perspective-based LLM Assessments in Situations with High-stakes), a meticulously curated dataset consisting of 345 high-impact dilemmas along with 3,795 individual perspectives of diverse values. CLASH enables the study of critical yet underexplored aspects of value-based decision-making processes, including understanding of decision ambivalence and psychological discomfort as well as capturing the temporal shifts of values in the perspectives of characters. By benchmarking 14 non-thinking and thinking models, we uncover several key findings. (1) Even strong proprietary models, such as GPT-5 and Claude-4-Sonnet, struggle with ambivalent decisions, achieving only 24.06 and 51.01 accuracy. (2) Although LLMs reasonably predict psychological discomfort, they do not adequately comprehend perspectives involving value shifts. (3) Cognitive behaviors that are effective in the math-solving and game strategy domains do not transfer to value reasoning. Instead, new failure patterns emerge, including early commitment and overcommitment. (4) The steerability of LLMs towards a given value is significantly correlated with their value preferences. (5) Finally, LLMs exhibit greater steerability when reasoning from a third-party perspective, although certain values (e.g., safety) benefit uniquely from first-person framing.

1 INTRODUCTION

CLASH addresses whether LLMs can make appropriate judgments in high-stakes dilemmas from diverse, contextualized perspectives. It reveals difficulties with ambivalence, value shifts, transferable reasoning strategies, and steerability.

Dataset contribution

  • CLASH introduces 345 human-written high-stakes dilemmas enriched with narrative perspectives representing diverse values.The dataset targets medical, legal, financial, and other consequential decision settings rather than everyday dilemmas.Source: §1 · pp. 1–2

Ambivalent decisions

  • 24.06 accuracy for GPT-5 and 51.01 for Claude-4-Sonnet show that even strong proprietary thinking models struggle with ambivalent decisions.CLASH explicitly incorporates ambivalent settings instead of restricting dilemmas to binary choices.Source: §1 · p. 2

Value shifts

  • A 42.95-point average accuracy drop under shifted values, reaching 66.43 points for GPT-4o-mini, exposes weak reasoning about updated perspectives.The benchmark examines temporal value changes by prompting models with shifted character values.Source: §1 · p. 2

Reasoning behavior

  • Backchaining and verification transfer poorly from math and game reasoning, while early commitment and overcommitment characterize value-reasoning failures.The analysis also suggests pragmatic ethics as a potentially useful framework for eliciting more effective reasoning.Source: §1 · p. 2

Steerability

  • Conditional steerability is negatively correlated with model preferences (r = −0.243, p < 0.005), and third-party framing generally improves steering.Some value pairs instead benefit more from first-person framing.Source: §1 · p. 2

2 DESIGN OF CLASH

CLASH represents high-stakes dilemmas through situations, actions, value rationales, and character descriptions, then tests judgments across static and dynamic value perspectives. Its questions separately assess acceptability, discomfort, and sensitivity to value shifts.

Elements in CLASH

  • CLASH combines high-stakes situations, controversial actions, supporting and opposing rationales, and character descriptions presented as Character A’s perspective.The situations contain rich contextual details intended to reflect real-life decision complexity.Source: §2 · pp. 2–3

Character Description Categories

  • Character descriptions comprise static categories—Straightforward, Simple Contrast, and Swayed Contrast—and dynamic categories—Shift, Half-Shift, and False-Shift.The design produces 11 descriptions per situation, with static categories holding values constant and dynamic categories varying them.Source: §2 · pp. 3–4

Static questions

  • Static evaluation asks whether an action is acceptable and whether performing or rejecting it would cause psychological discomfort.Acceptability uses Yes, No, or Ambiguous, while discomfort uses Yes or No.Source: §2 · p. 3
  • The label “Ambiguous” marks indeterminate acceptability when competing character values prevent a definitive judgment, rather than directly naming first-person ambivalence.The distinction reflects the third-person framing used in the main experiments.Source: §2 · p. 3

Dynamic questions

  • Dynamic evaluation asks separately about previous and current value preferences, using ternary judgments to test sensitivity to value shifts.Shift, Half-Shift, and False-Shift respectively encode changed, partially changed, and stable preferences under challenging context.Source: §2 · p. 4

3 DATASET CONSTRUCTION PROCEDURE FOR CLASH

CLASH is constructed through four stages that collect and verify dilemmas, generate and refine value rationales and character descriptions, and validate intended responses. The resulting benchmark contains 345 situations and 3,795 value perspectives.

Benchmark curation process

  • The curation pipeline has four stages: situation and action generation, rationale generation, character-description generation, and dataset validation.The stages combine automated generation with human inspection or refinement.Source: §3 · p. 4

Step 1) Situation Collection and Action Generation

  • GPT-4o generates actions from collected dilemmas, while authors manually inspect whether the actions align with the situations.The dilemmas are collected from websites featuring high-impact domains.Source: §3 · p. 4

Step 2) Value-Related Rationales Generation

  • Eight native English-speaking students evaluate and refine GPT-4o-generated supporting and opposing rationales for relevance, clarity, and formatting.A checklist guides quality control for each rationale.Source: §3 · p. 4

Step 3) Character Description Generation

  • Five inspectors review GPT-4o-generated character descriptions to ensure alignment with the situations and intended categories.Descriptions are generated using prompts tailored to each category.Source: §3 · p. 4

Step 4) Dataset Validation

  • 345 situations yield 3,795 individual perspectives, while validation achieves an average pairwise Cohen’s Kappa of 0.985.Only 21% of events were judged US-focused, with moderate cross-annotator agreement of 0.458.Source: §3 · p. 4

4 EXPERIMENTS AND ANALYSES

CLASH evaluates LLM judgment across ambivalence, discomfort, value shifts, reasoning chains, conditional steerability, and question framing in high-stakes dilemmas. Models perform well on discomfort but struggle with ambivalence and shifted values, while value reasoning exhibits distinct failure patterns and framing-dependent steerability.

4.1 MAIN RESULTS

  • Claude-4 Sonnet achieved the highest overall accuracy at 88.89, while models generally performed poorly on ambivalence and current value-shift questions.RLMs outperformed their LLM counterparts within the Qwen, GPT, and Claude families, but produced substantially longer outputs on average.Source: §4.1 · p. 5, Table 1 · p. 5
  • Models struggled to distinguish clear-cut from ambivalent decisions, with Mistral-123B reaching only 62.90 accuracy in Simple Contrast.GPT-5 and Deepseek-3.1 performed best on Swayed Contrast, whereas open-weight Mistral models led Simple Contrast; Qwen3-32B reasoning also paired hedging with definitive answers.Source: §4.1 · p. 6, Figure 3 · p. 6
  • GPT-4o reached 96.50 accuracy on Straightforward discomfort and GPT-5 reached 98.61 on Swayed Contrast discomfort.Straightforward characters’ extreme perspectives were associated with preferring action over inaction.Source: §4.1 · p. 6
  • Value-shift accuracy fell by an average of 42.95 points, with GPT-4o-mini showing the largest drop of 66.43 points on current questions.The current question was harder than the previous question across all models, with p < 0.0001.Source: Table 1 · p. 5, §4.1 · p. 6

4.2 FURTHER REASONING CHAIN ANALYSES

  • Successful value reasoning used less backward chaining and verification than math and game reasoning, while unsuccessful chains showed early commitment and overcommitment.Successful chains also emphasized pragmatic and rights-based ethics more than unsuccessful chains, with pragmatic ethics especially prominent.Source: §4.2 · p. 7, Figure 4 · p. 7

4.3 CONDITIONAL STEERABILITY ANALYSIS USING DIVERSE VALUE FRAMEWORKS

  • Conditional steerability was negatively correlated with inherent value preference (r = −0.243, p < 0.005), and larger models showed higher overall steerability.Smaller models were more likely to steer unevenly toward one value over the other.Source: §4.3 · pp. 8–9

4.4 QUESTION FRAMING COMPARISONS VIA FIRST- VS. THIRD-PERSON PERSPECTIVES

  • Third-person framing generally increased steerability, although first-person framing was particularly effective for steering toward Safety relative to Self-Esteem.Overall performance declined significantly under first-person framing, with p < 0.001.Source: §4.4 · p. 9

5 RELATED WORK

Prior value-reasoning benchmarks often use fixed ground truths, human judgments, indirect plurality, or everyday dilemmas. CLASH extends this work with narrated, contextualized perspectives and high-stakes value conflicts, including conditional steerability between competing values.

5 RELATED WORK

  • Existing moral benchmarks compare outputs with predefined or human ground truths, whereas other studies examine preferences in ambiguous dilemmas without a clear reference decision.CLASH bridges these approaches by enabling verifiable evaluation when no clear-cut ground-truth choice exists.Source: §5 · p. 9
  • Few benchmarks represent human values through narrated perspectives; metadata, disagreement rates, and demographic proxies miss relational trade-offs expressed in narratives.CLASH incorporates these interactions through character descriptions to better reflect the complexity of human value perspectives.Source: §5 · p. 9
  • Existing datasets focus mainly on everyday scenarios from forums, crowdsourced situations, or synthetic generation, whereas CLASH selects expert-authored high-stakes conflicts.High-stakes dilemmas require domain expertise and heightened sensitivity to value conflicts.Source: §5 · p. 9
  • Prior value steering often treats values as independently desirable or reinforces one side without modeling conflicts and trade-offs.CLASH’s conditional steerability asks whether a model can follow value X when it conflicts with value Y.Source: §5 · pp. 9–10

6 CONCLUSION

CLASH shows that frontier LLMs still struggle with ambivalence, discomfort, and value shifts in high-stakes dilemmas. Its analyses identify value-specific reasoning failures and framing-dependent steerability, motivating further work on human-aligned value reasoning.

6 CONCLUSION

  • Frontier LLMs struggle with decision ambivalence, psychological discomfort, and value shifts, revealing limitations in high-stakes value judgment.Source: §6 · p. 10
  • Value reasoning produces early commitment and overcommitment rather than simply transferring successful math or game reasoning behaviors.Source: §6 · p. 10
  • Conditional steerability is negatively correlated with value preference, while third-person framing is generally more effective.Source: §6 · p. 10

ETHICS STATEMENT

The paper frames CLASH as a methodological and diagnostic benchmark for high-stakes value-sensitive reasoning, not evidence that models are ready for autonomous deployment. It also documents dataset construction safeguards and supporting materials for evaluating rich, real-world dilemmas.

ETHICS STATEMENT

  • CLASH scores alone should not guide deployment in high-stakes domains, which require domain-specific evaluation, human oversight, and legal and ethical alignment.Source: ETHICS STATEMENT · p. 10
  • The dataset uses LLM-generated actions, rationales, and character descriptions that human inspectors and annotators inspect, edit, and validate.Source: ETHICS STATEMENT · pp. 10–11

APPENDIX

  • The appendix provides dataset-design, construction, experimental, supplementary-results, reasoning-chain, and conditional-steerability details, including prompts and examples.Source: APPENDIX · p. 15

A.1 FORMULATION OF THE QUESTION

  • Questions are tailored to whether an action was carried out, considered, plausibly under consideration, or questioned within the narrative.Source: §A.1 · p. 15

A.2 IMPORTANCE OF THE DETAILS IN SITUATIONS

  • Rich situation details are retained because they capture real-world complexity, and performance differs significantly between original and summarized narratives.The comparison uses repeated evaluations, a two-tailed Wilcoxon test, and average differences across general, discomfort, and temporal questions.Source: §A.2 · p. 15

A.3 FULL DATASET EXAMPLE

A.4 DATASET COMPOSITION: TOPICS AND VALUE FRAMEWORKS

CLASH spans four high-stakes domains and represents value dimensions across several frameworks. Its construction combines filtered web-sourced situations with prompted generation and multiple rounds of checklist-based human inspection.

A.4 DATASET COMPOSITION: TOPICS AND VALUE FRAMEWORKS

  • Situations cover Journalism/Media (33.33%), Business (31.88%), Medical (23.48%), and Government/Politics (11.30%).Source: §A.4 · p. 18
  • Value coverage is broad but uneven: World Values Survey dimensions each exceed 75%, while Aristotle’s Modesty and Patience appear in only 0.58% and 0.29% of dilemmas.Source: §A.4 · p. 18

B.1 SITUATION COLLECTION AND ACTION GENERATION

  • Situations are collected from public websites, screened for contamination and narrative suitability, and paired with actions generated by GPT-4o before manual inspection.Source: §B.1 · p. 19, Table 3 · p. 19

B.2 VALUE-RELATED RATIONALE GENERATION

  • GPT-4o generates value-related rationales that eight native English-speaking students refine using a checklist covering clarity, formatting, and relevance.Source: §B.2 · p. 19

B.3 CHARACTER DESCRIPTION GENERATION

B.4 DATASET VALIDATION PROCESS

The validation process checks whether character descriptions elicit their intended perspective-based answers and uses independent annotation to resolve disagreements. Agreement is very high, supporting the reliability of the labels.

B.4 DATASET VALIDATION PROCESS

  • An initial annotator evaluates descriptions without knowing the predefined categories, and a second annotator independently breaks ties when answers disagree with intended labels.Source: §B.4 · p. 25
  • Average pairwise Cohen’s Kappa across five inspectors and three annotators is 0.985.Source: §B.4 · p. 25
  • Explicit character perspectives make intended value trade-offs clearer than questions posed without a specified perspective.For example, a character prioritizing fairness selects the person who waited longer rather than the person with higher survival chances.Source: §B.4 · p. 25

C EXPERIMENTAL SETUP

The experiments evaluate multiple language and reasoning models with category-specific prompts for static and dynamic perspectives. Full results cover static and dynamic categories, with human accuracy reported on a sampled subset.

C.1 MODELS

C.2 PROMPTS FOR GENERATING ANSWERS TO THE QUESTIONS

  • Static categories use General and Discomfort questions, whereas dynamic categories use Previous and Current questions.Source: §C.2 · p. 26

D.1 FULL EVALUATION RESULTS

  • Full static and dynamic evaluation results are reported in Tables 5 and 6, alongside average accuracy from two context-unaware human annotators on 50 randomly sampled situations.Source: §D.1 · p. 26

D.2 ERROR ANALYSIS

The analysis identifies recurring reasoning failures, including value confusion, situational misinterpretation, early commitment, and overcommitment, while examining discomfort, steerability, framing, value mapping, and preference calculations.

D.2 ERROR ANALYSIS

  • Smaller models often misinterpreted scenario details or produced inconsistent final answers despite reasoning toward the correct rationale.Larger Llama-3.3-72B did not exhibit the value-confusion and situational-rigidity deficiencies observed in Llama-3.1-8B-Instruct.Source: §D.2 · p. 29
  • Approximately 85% of incorrect responses involved value confusion, making it the most prevalent observed error.Models ambiguously integrated conflicting values instead of enforcing the intended prioritization.Source: §D.2 · pp. 26, 29

D.3 PERFORMANCE DROP IN DISCOMFORT SETTING

  • Negating actions reverses the discomfort-prediction pattern in Straightforward cases but not in Swayed Contrast cases.Models become more likely to correctly predict no discomfort when performing, rather than refraining from, the action in Straightforward settings.Source: §D.3 · p. 32

D.4 STEERABILITY RESULTS USING ALL PAIRS

  • When all eligible value pairs are included, the base-preference–steerability correlation becomes nonsignificant at 0.046 (p = 3.64e−1).Smaller models remain more likely to favor one value, whereas larger models retain higher overall steerability.Source: §D.4 · p. 32

E.1 REASONING OUTPUTS

  • Examples contrast successful and unsuccessful ambivalent reasoning chains across Claude-4 Sonnet, Qwen3-32B, and Deepseek-3.1.The examples specifically expose hedging expressions in model-generated reasoning.Source: §E.1 · p. 33, Figure 21 · p. 34, Figure 22 · p. 34

E.2 PROMPTS FOR COGNITIVE CHARACTERISTICS

  • The analysis probes backward chaining, verification, early commitment, and overcommitment with dedicated prompts across ambivalence and discomfort settings.Examples show Qwen3-32B committing quickly and then reinforcing its choice or shifting prematurely from discomfort identification to action selection.Source: §E.2 · p. 33, Figure 31 · p. 37, Figure 32 · p. 37

F.4 DETAILED VALUE MAPPING PROCESS

  • Value mapping combines entailment-model and GPT-4o judgments, filtering, and human review to associate rationales with intermediate values and broader value frameworks.Human evaluators judged 98.77% of mappings relevant and 78.00% comprehensive, with agreement scores of 0.823 and 0.471.Source: §F.4 · p. 41

F.5 PREFERENCE AND STEERABILITY CALCULATION

  • Steerability is computed from preference ratios for competing values, while full comparisons report differences between steering toward each side of a value pair.The calculation uses b/a for Safety and c/d for Self-Esteem, and per-pair bars encode the right-minus-left steerability difference.Source: §F.5 · p. 42

F.6 PER VALUE-PAIR STEERABILITY COMPARISON RESULTS

Per-value-pair steerability results compare the two directions of steering for each competing value pair and model.

F.6 PER VALUE-PAIR STEERABILITY COMPARISON RESULTS

  • Each bar shows the right value’s steerability minus the left value’s steerability, so its direction identifies which competing value is more steerable.Leftward bars indicate greater steerability toward the left value; rightward bars indicate the reverse.Source: §F.6 · p. 43, Figure 40 · p. 43

Related papers

Loading 2504.10823v4…